产品发布/更新
-
OpenRouter MCP server released
OpenRouter launched an MCP server that provides coding agents with real-time model data, benchmark rankings, pricing,…
-
Midjourney previews V8.2 and speeds up draft mode
Midjourney has introduced two updates. First, a `--preview` parameter lets users experience V8.2's aesthetics and…
-
Codex officially available in the ChatGPT mobile app
OpenAI announced that Codex is now generally available (GA) in the ChatGPT mobile app, with new one-to-one device…
-
Runway releases Agent 2.0
Runway has released Agent 2.0 to help marketers create, test and optimize ads, videos and marketing campaigns. Brand…
-
Midjourney V8.1 草稿模式新增随机风格功能
Midjourney V8.1 的草稿模式(draft mode)添加了随机风格功能。用户在提示词中加入 `--sref random` 即可一键生成 24 张不同风格的图片。开启草稿模式可通过点击提示栏的 ⚡ 图标或添加 `--draf…
-
Claude Code v2.1.193 released
Claude Code v2.1.193 adds the `autoMode.classifyAllShell` setting, routing all Bash/PowerShell commands through the…
-
Google Finance launches new Android app and portfolio features
This week, Google Finance launched its official Android app and simultaneously rolled out global portfolio tracking.…
-
Interactive Brokers integrates with Grok: portfolio analysis, scenario modeling, and real-time trading instructions
Interactive Brokers recently integrated with Grok, allowing users to link existing accounts for free in minutes…
行业动态
-
美国政府要求OpenAI暂缓GPT-5.6广泛发布
The Information 报道,美国政府因安全顾虑要求 OpenAI 暂缓 GPT-5.6 的广泛发布,改为推出受控预览版。OpenAI 计划先向小部分合作伙伴提供早期访问,并由政府逐客户审批准入。这一非常规做法主要担忧模型在自动化…
-
Meta employees warn AI content moderation deployed too fast
In 2025, Meta replaced about half of human moderation requests with large language models and plans to raise the AI…
-
General Intuition raises $320M to train general AI agents with game data
General Intuition has raised $320 million at a $2.3 billion valuation, bringing total disclosed funding to $454…
论文研究
-
OpenAI internal report: how agent Codex is changing work
OpenAI observed between August 2025 and June 2026 that agent product Codex replaced ChatGPT as the primary work tool,…
-
用线性弹性缓存优化云经济
Google Research 与 Google Cloud 提出线性弹性缓存,将缓存管理转为线性成本优化问题,动态调整大小以最小化总拥有成本。为每条数据引入“滑雪租赁”决策框架,在租用内存(持续付费)与购买缺失(缓存未命中惩罚)间选择,…
技巧与观点
-
@exponentialview releases State of the AI Economy report: AI economy annualized revenue exceeds $175 billion
Based on deduplicated consumer AI spending statistics, the report finds actual AI revenue over the past 12 months…
-
Most mainstream AI chatbots lean left politically, including "anti-woke" models
A Washington Post investigation shows most mainstream AI chatbots are clearly left-leaning on political issues. OpenAI…
-
OLMo Hybrid vs Transformer: Hybrid models show clear advantage on content words but little on repeated phrases
By comparing the 7B-parameter OLMo 3 (Transformer) with OLMo Hybrid (hybrid architecture), experiments found that the…
-
Meta 隐私感知基础设施的资产分类:混合模式将 LLM 蒸馏为确定性规则
Meta 在 Privacy-Aware Infrastructure (PAI) 的资产分类中采用混合模式:先构建含代码、血缘、语义标注的上下文证据,再调用 LLM 处理歧义、冷启动和新颖资产;人工审核标签与模型推荐严格隔离。LLM 不…
-
跨模型与任务的 GitHub Copilot agentic harness 性能与效率评估
GitHub Copilot agentic harness 在多个基准测试中表现强劲,同时具备领先的 token 效率,并支持在 20 多个模型间灵活选择。
-
Sail Research 构建集群感知编排,加速异步推理
推理市场是软件中最大的市场。AI工作负载正从同步聊天转向异步、多轮智能体,运行时长可达数小时。Sail Research 为此构建了集群感知(fleet‑aware)编排系统,以最大化每美元推理支出的吞吐量。
-
Generative AI Fizzle™:生成式AI泡沫正在缓慢消退
Gary Marcus 昨日提出新术语 Generative AI Fizzle™,认为生成式AI行业估值过高,投资者对 hype 与利润的落差失去热情。LLM 已商品化,价格战激烈,提供商盈利艰难。昨日一款新的中国开源模型发布,可能进一…